Predictive Churn Scoring

SaaS Tracker trains an XGBoost model nightly on your workspace’s event data to assign every active user a churn probability score (0–1).

How it works

1. Feature extraction — the ML service computes behavioral features per user over the past 30 days: event frequency, recency, feature breadth, engagement trend, session count
2. Model training — XGBoost is retrained nightly if ≥ 500 users have been active in the past 90 days; otherwise the previous model is reused
3. Scoring — each active user receives a churnProb score; users not seen in 30 days are excluded
4. Dashboard — scores appear in the Users table and the Churn KPI card on the Overview page

Score interpretation

Triggering actions on churn score

You can use churn score as an audience filter in Messages:

User filter: churnScore > 0.5

This lets you send targeted in-app messages to high-risk users without any manual segmentation.

Model quality

The model’s AUC is monitored by Grafana. If AUC drops below 0.65, an alert fires — meaning the model is no better than random chance and manual review is needed.

Limitations

• Scores are computed nightly, not in real time
• The model performs best with ≥ 1 000 monthly active users
• Cold-start workspaces (< 90 days of data) use a heuristic recency model until enough data accumulates